Initial K3 snapshot: 0.5B KDA/MLA/MoE train path

Standalone tree split from LLMRL/projects/kda. Includes Triton dt_bias
backward fix, train_k3 --preset 0.5b, SFT, Docker runtime, and tests.

tests/integration/test_schedule.py

@@ -0,0 +1,21 @@
+from kda.training.schedule import lr_scale, total_opt_steps
+
+
+def test_warmup_then_cosine_floor():
+    assert abs(lr_scale(0, warmup=10, total_opt=100) - 0.1) < 1e-9
+    assert abs(lr_scale(9, warmup=10, total_opt=100) - 1.0) < 1e-9
+    assert abs(lr_scale(10, warmup=10, total_opt=100) - 1.0) < 1e-6
+    end = lr_scale(99, warmup=10, total_opt=100)
+    assert abs(end - 0.1) < 1e-6
+
+
+def test_horizon_prefers_the_earlier_stop():
+    # 8.2M tokens @ batch 2 seq 2048 acc 8 -> 250 opt
+    opt_from_tokens = total_opt_steps(
+        max_tokens=8_192_000, max_micro=10_000, batch=2, seq_len=2048, grad_acc=8
+    )
+    assert opt_from_tokens == 250
+    opt_from_micro = total_opt_steps(
+        max_tokens=10**12, max_micro=2000, batch=2, seq_len=2048, grad_acc=8
+    )
+    assert opt_from_micro == 250